Master Claude

Frontier-grade discipline for any Claude model. Master Claude is a Claude Code plugin that makes Opus, Sonnet, or Haiku work the way Anthropic's frontier tier works: question the request before building, brainstorm adversarially, decide with precision, verify against current sources (because laws, tax rates, and APIs change month to month), execute with evidence — and remember everything across sessions through an automatic, Obsidian-compatible mind-map vault.

Works across Claude Code (CLI + desktop), Cowork, claude.ai Chat, and the Claude API/Agent SDK. Nothing in it is user-specific — install it for yourself or your whole team.

Quick start

npx master-claude

That's it. The installer auto-detects Claude Code on macOS, Windows, or Linux (CLI and desktop app both), registers the plugin, and installs it. Then open Claude Code and run:

/mc-setup

— a five-minute wizard that sets up your mind-map vault, company book, operating doctrine, and personal defaults. From then on, everything below just happens.

Installer options

| Command | What it does | |---|---| | npx master-claude | detect + install (safe to re-run; it updates) | | npx master-claude --with-doctrine | also append the Fable Doctrine to ~/.claude/CLAUDE.md | | npx master-claude --skills-only | no plugin system? copy the 12 skills to ~/.claude/skills | | npx master-claude --dry-run | preview every step, change nothing | | npx master-claude --uninstall | remove the plugin, marketplace, and staged files |

What you get

• 🧠 A reasoning pipeline on every substantial task — question → brainstorm → decide → ground → plan → execute → verify → write back, with depth that scales to stakes (small asks stay fast).
• 🗺️ An automatic mind-map memory — every project gets an Obsidian-compatible vault of atomic, [[wikilinked]] notes. Sessions load only the index and expand just the branch they need; new sessions resume instead of rediscovering. Related vaults from other projects are detected and offered for linking — with your consent, never silently.
• 🔍 Nothing stale ships — laws, tax rates, SDK signatures, model pricing: anything fast-moving is verified against current official sources before use, cited with a verified-on date.
• ⚖️ Indian legal studio — draft/review/redline under the Companies Act 2013, Contract Act 1872, BNS/BNSS/BSA, and state-wise stamp/S&E/PT rules.
• 💰 Indian finance studio — Ind AS/ICAI accounting, GST/TDS, Income-tax Act 2025, ledgers, and formula-driven Excel models.
• 🏢 Company Book — your CIN, GSTINs, addresses, signatories, logo, and brand, gathered once (with format validation) and pulled into every deck, contract, invoice, and video automatically.
• 📊 PPT studio & 🎬 Video studio — branded presentations (PPTX/HTML) and videos (Remotion/HeyGen, ElevenLabs/Sarvam/offline voice, captions), with approval gates before the expensive steps.
• 🏗️ Platform builder — discovery → stack recommendation → element-level feature breakdown → clickable design previews → phased, checkpointed execution.
• 🔐 Secrets that never touch the model — a collector script (.sh + .ps1) takes credentials as hidden terminal input, and an enforced env-guard hook blocks the model from ever reading .env.
• 💾 Crash-proof progress — checkpoints survive crashes, restarts, and context compaction. Interrupted work resumes mid-phase.

Commands

Skills trigger automatically from plain language; commands make invocation deterministic and skip the routing.

| Command | Does | |---|---| | /mc-setup | first-run wizard: vault, company book, doctrine, defaults, connectors | | /legal review <file> · /legal draft <type> | Indian legal work, mode pre-selected | | /finance gst <q> · /finance model <what> | Indian finance, tax, and modeling | | /deck <topic> | branded presentation pipeline | | /video <brief> | branded video pipeline | | /build <idea> | platform builder, all eight stages | | /book | create / check / update the company book | | /checkpoint <msg> | save a resume point | | /mc-help [topic] | guide to everything, or one focused answer |

The mind-map (how memory works)

vault/
├── HOME.md          ← the map: one line per note, [[wikilinks]] as edges
├── knowledge/       ← durable facts, gotchas, constraints
├── decisions/       ← decision records with confidence + revisit triggers
├── references/      ← external sources, each with a verified-on date
├── sessions/        ← short logs: outcome, dead ends, next step
└── mocs/            ← sub-maps for areas that outgrow HOME.md

• Automatic: in project directories the vault is created on session start; in chat, Claude starts one the first time a session produces something durable. You never ask for memory — you have it.
• Selective: only HOME.md loads by default. Claude follows links like expanding one branch of a mind map — the rest stays folded, saving your context window for actual work.
• Cross-project: vaults register in ~/.master-claude/vaults.json. When today's work overlaps another project's vault, Claude proposes linking it — one line, your call. Agreed links live under ## Linked vaults and let sessions draw on the other project's notes without merging anything.
• Yours: open vault/ in Obsidian — graph view, backlinks, and editing all work natively. Same files, no sync layer.

Safety model

• Secrets: run the collector (collect-secrets.sh / collect-secrets.ps1) — you type values with echo off, the script writes .env directly, the model only ever learns key names. The env-guard hook then blocks any attempt to read, write, grep, or cat secret env files (placeholders like .env.example stay usable). Never paste a secret into chat; the environment is built so you never need to.
• Professional boundaries: the legal and finance studios prepare workings and drafts, verify citations, and tell you when something needs a licensed advocate/CA/CS — statutory filings, court matters, and audit opinions always do.
• Fail-open guards: every hook exits cleanly on error — the safety layer can slow a mistake, never your session.

Install on each surface

Claude Code — CLI & desktop (recommended): npx master-claude (above). Manual alternative:

claude plugin marketplace add ujjwalsittu/master-claude
claude plugin install master-claude@master-claude

claude.ai Chat: Settings → Capabilities → Skills → upload the zips from dist/. Create a Project and paste FABLE-DOCTRINE.md into its instructions; add your vault/ to Project knowledge (or sync via a GitHub/Drive/Notion connector).

Cowork: consumes the same plugin — install once, it's there.

API / Agent SDK: system prompt ← FABLE-DOCTRINE.md; skills ← plugins/master-claude/skills/*; agents ← plugins/master-claude/agents/*; mount vault/ in the workspace.

Configuration

| Where | What | |---|---| | vault/config.md | your defaults (state, ₹ formats, FY, video language, stack preference) — created by /mc-setup, edit anytime | | company-book/ | identity + brand; /book manages it, every deliverable reads it | | ~/.claude/CLAUDE.md | the Fable Doctrine (system-level operating laws) | | .fable/ | per-project state: checkpoints, change log, activity (gitignore it) | | ~/.master-claude/ | staged marketplace + cross-project vault registry |

Updating · Uninstalling

npx master-claude              # re-run = update
npx master-claude --uninstall  # clean removal

Plugin users on the GitHub marketplace: /plugin marketplace update master-claude.

Troubleshooting

• A skill didn't trigger → invoke it directly: /legal, /deck, or "use the ppt-studio skill". Commands are deterministic.
• Hook didn't fire / no vault appeared → hooks need the plugin install (not --skills-only), and auto-creation only happens in project-like directories (.git, package.json, …). Run /mc-setup to create one anywhere.
• env-guard blocked something legitimate → it points to the sanctioned path (collect-secrets, .env.example). If you truly need an exception, that's a conscious user action — do it outside Claude.
• Windows → the installer and env-guard are native; hook shell snippets and collect-secrets.sh want Git Bash (ships with Git for Windows), or use collect-secrets.ps1 in PowerShell.
• A legal/tax/API answer looks stale → say "verify that" — it forces the grounded-research pass with current notifications.
• Anything else → /mc-help <your question>.

For maintainers (publishing this package)

# 1. npm (powers npx master-claude)
npm login
npm publish                     # bump version in package.json first

# 2. GitHub (powers the plugin marketplace)
git push origin main            # users: /plugin marketplace add ujjwalsittu/master-claude

Keep the version synced in three files: package.json, plugins/master-claude/.claude-plugin/plugin.json, and .claude-plugin/marketplace.json (metadata.version).

Anthropic's official plugin directory (pre-installed in every Claude Code; installs as master-claude@claude-plugins-official): submit the public GitHub repo via the plugin directory submission form. Requirements this repo already meets: .claude-plugin/plugin.json, LICENSE inside the plugin, README, homepage. Note the name slug is immutable once published.

Regenerate the claude.ai skill zips after editing any skill: ./package.sh.

Contributing

The structure is deliberately templatable — a new domain studio is one folder: plugins/master-claude/skills/<name>/SKILL.md (frontmatter name + a trigger-rich description), optional references/ for depth, scripts/ for deterministic helpers, and a "Fable discipline" section tying it into the pipeline. PRs welcome.

License

MIT © 2026 Ujjwal Sittu